Inferential Wasserstein Generative Adversarial Networks

Abstract Generative adversarial networks (GANs) have been impactful on many problems and applications but suffer from unstable training. The Wasserstein GAN (WGAN) leverages the distance to avoid caveats in minmax two-player training of GANs has other defects such as mode collapse lack metric detect convergence. We introduce a novel inferential (iWGAN) model, which is principled framework fuse autoencoders WGANs. iWGAN model jointly learns an encoder network generator motivated by iterative primal-dual optimization process. maps observed samples latent space data space. establish generalization error bound theoretically justify its performance. further provide rigorous probabilistic interpretation our under maximum likelihood estimation. iWGAN, with clear stopping criteria, advantages over autoencoder GANs. empirical experiments show that greatly mitigates symptom collapse, speeds up convergence, able measurement quality check for each individual sample. illustrate ability obtaining competitive stable performances benchmark datasets.